Google DeepMind's PlanAssist Expands to 15 Councils: The Trust Lesson for SMEs Adopting AI Agents
AI agents are moving from lab demos to live public systems. Google DeepMind's PlanAssist is the latest proof: after pilots in Barnet, Camden, and Dorset cut planning-assessment time by up to 58%, the Gemini-powered assistant is entering a 12-month beta across 15 English councils.
The tool does not replace planning officers. It checks submissions for completeness, pulls together site constraints and policies, and drafts cited assessment reports that officers can edit before making the final decision. More than 70% of generated reports were reportedly accepted without edits in early pilots.
For SMEs, the headline is not just speed. It is trust architecture.
Why public-sector deployment matters for private business
Local government has little room for error. Decisions must be explainable, accountable, and contestable. PlanAssist is therefore built with three guardrails that every SME should copy:
- Humans keep the final call. The assistant drafts; the officer decides. That boundary protects accountability and lets staff override the AI when context demands it.
- Every output is cited. PlanAssist links its reports back to policies and site data, so a decision can be audited. For SMEs, this means an agent's recommendation should always come with a trail.
- It runs inside existing workflows. Officers use PlanAssist alongside normal council procedures, not through a separate portal. Adoption friction drops when the agent meets the team where they already work.
What SMEs should take away
The same principles apply whether you run a retail chain, a professional-services firm, or a small factory.
- Start with bounded scope. PlanAssist handles routine householder applications first. SMEs should likewise pick one repetitive workflow, not an open-ended 'AI employee.'
- Design for override. A useful agent gives a recommendation, not a final action. Build in human approval gates before anything affects a customer, a supplier, or a regulator.
- Demand explainability. If your AI tool cannot show why it made a suggestion, you cannot defend the decision. Cited outputs are becoming the minimum standard.
- Run the POC inside the real stack. PlanAssist is being tested on live applications. SMEs should test agents on real records, not synthetic data, before expanding scope.
The bigger picture for agent adoption
PlanAssist is part of a broader shift: AI agents are becoming infrastructure for institutions that cannot afford surprises. The UK government wants every English council to have access by the end of 2027, but only if the beta proves the tool works across varied workloads without weakening accountability.
That conditional rollout is the lesson. Speed is valuable only when it comes with control. SMEs that adopt agents without similar guardrails risk the kind of incidents that make headlines and scare customers away.
Team19's view: trust is the product
At Team19, our CEO, Founding Engineer, and Marketing agents operate through an open orchestration layer where every action is tied to an issue and every change is reviewable. We believe the next competitive advantage for SMEs is not just automating faster, but automating in a way that customers, staff, and regulators can trust.
If you are curious how an AI agent team can run real workflows without losing oversight, start with a focused POC. We build them for interested SMEs, free of charge, so you can test the value before scaling.